[
https://issues.apache.org/jira/browse/HDFS-5999?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13909686#comment-13909686
]
Tsz Wo (Nicholas), SZE commented on HDFS-5999:
----------------------------------------------
Patch looks good. A minor suggestion: Let's call hasRollbackFSImage in
FSEditLogLoader. Then, we could keep the saveNamespace parameter in
startRollingUpgradeInternal. Passing txid is harder to understand.
> Do not create rollback fsimage when it already exists
> -----------------------------------------------------
>
> Key: HDFS-5999
> URL: https://issues.apache.org/jira/browse/HDFS-5999
> Project: Hadoop HDFS
> Issue Type: Sub-task
> Components: datanode, ha, hdfs-client, namenode
> Reporter: Jing Zhao
> Assignee: Jing Zhao
> Attachments: HDFS-5999.000.patch
>
>
> Currently in FSEditLogLoader#applyEditLogOp, if the op is
> OP_START_LOG_SEGMENT, we do an extra rollback fsimage with condition
> "totalEdits > 1".
> This may cause some issue when we restart a NN with new software. If there is
> no extra checkpoint after the rollback image, the NN may generate a new
> rollback image using the new software, and replace the old rollback image.
> Then later if we want to do rollback the NN, which uses the old software, may
> not be able to load the rollback image.
--
This message was sent by Atlassian JIRA
(v6.1.5#6160)
